t2m: retrain the generative model on v5 bind-referenced canonical data

[fernandotonon]

Part of epic #837. The current t2m CVAE (v4) was trained on canonical extractions whose absolute world quats mixed per-rig conventions (bind-frame conjugation, pre-#843) — the model learned a convention mush, which is why its output is a gentle average and why it emits no usable reference triple (its clips need the synthetic standing pose from PR #843's model-path fix).

Plan

• Data: re-extract the full corpus with the calm-frame extractor (PR feat(anim): canonical clip extraction + motion-library v5 builder (t2m-v2 Slice B, #839) #843) — --dump-canonical now yields cross-rig-consistent world quats + per-clip restWorld/restDir. Rebuild the training windows from those dumps (extend scripts/prep-t2m-v4.py → v5), keeping the v4 lessons: 30fps world-frame windows with neutral starts, absolute-pose decoder, derived-local supervision, z=0 latent-dropout supervision.
• Reference triple for model output: since training data is now convention-consistent, export a canonical restWorld/restDir alongside the model (dataset mean at the reference frame) and emit it from MotionGenerator — model clips then ride the SAME bind-referenced direction retarget as templates, retiring the synthetic-standing-pose shim.
• Eval gate: harness parity vs the template path on walk/run/dance (joint deltas + fox renders); ship only if the model closes on the template quality bar, else it stays experimental/opt-in.
• Bump the model files on the HF repo (motion/t2m.onnx + vocab) with backward-compatible naming or a version suffix.

Acceptance

• Model-generated walk on the Hip Hop fox is upright, bind-referenced, and visually comparable to the template take.
• Harness numbers recorded in the PR; templates remain the default until the gate passes.